Home → Insights → Food and drug enforcement data

China Food and Drug Enforcement: A Market Regulation Case Law Dataset

For a legal AI team building China regulatory coverage, a search for “food safety cases” answers several different questions at once. It may find a consumer seeking compensation, a criminal prosecution, a business challenging a regulator’s penalty, or an agency asking a court to enforce its decision. The product category is shared; the legal task is not.

A useful China food and drug enforcement case law dataset therefore starts with a narrower definition: court records concerning a regulator’s action, the action challenged, and the judicial procedure in which it was considered. It should support questions about sampling evidence, product classification, licensing, seizure, penalties and reconsideration without silently merging those questions into consumer litigation.

This guide sets out the boundaries for enterprise buyers, identifies official examples, and reports three live corpus queries checked on October 8, 2026. The counts illustrate retrieval choices rather than the size of an audited topic dataset. They are intended to help a Head of Data specify a pilot before selecting a production corpus.

Define the regulatory track before buying the data

The unit you need depends on what your assistant is meant to answer. An enforcement research tool might retrieve judicial discussion of an inspection or penalty. A consumer product needs private claims. An investigations tool may need criminal proceedings. Treating all three as one category produces plausible answers supported by the wrong kind of record.

These are dataset design distinctions. Related proceedings may concern the same goods, but each document still needs its own procedural label. Our product liability and recall dataset guide provides a broader overview of civil liability, consumer disputes, and administrative recalls; the present page focuses on retrieving and auditing regulatory court records. For the broader category, see the administrative litigation dataset overview.

This article is for informational purposes only and is not legal advice. The cited rules and judgments identify dated sources; verify the applicable law and facts before relying on them in a specific matter.

Official sources show what the category actually contains

The Supreme People’s Court’s April 28, 2022 publication of drug safety typical cases includes both criminal cases and administrative challenges. Its administrative examples include disputes involving drug manufacturers, penalties and reconsideration decisions. That official collection is a useful starting point for defining relevant questions, but it is a selected set of case summaries rather than a census of litigation. Source: SPC drug safety typical cases, 2022.

In Case Five, a manufacturer challenged a penalty involving production of an active pharmaceutical ingredient through an unapproved process. The published account describes rejection of its claims at first instance, dismissal of the appeal and rejection of a retrial application. For a dataset, those are distinct procedural events; counting the documents as unrelated controversies would obscure the history of one dispute. Source: SPC, Case Five.

The consolidated market supervision penalty procedure rules published on SAMR’s official site also identify useful research dimensions. Articles 31–32 address sampling and specialized testing; Articles 57–58 address notice and representations; Article 62 specifies contents of a penalty decision. The version on that page records amendments through September 2022. A historical case must be read against the rules applicable at the relevant time. Source: SAMR penalty procedure rules.

These sources help a buyer construct review questions: Was sampling disputed? Which standard was cited? What document records notice? They do not establish that a commercial corpus contains every underlying inspection record, laboratory report or administrative decision. That must be checked against the proposed delivery.

Three corpus queries, three different boundaries

On October 8, 2026, we ran the following queries against the Chinese court record index underlying this coverage check. The two agency queries require either of the stored administrative case-type labels, 行政 or 行政案件. They search the document body with the union of default phrase analysis and explicit ik_max_word phrase analysis. Exact total-hit tracking was enabled; all shards completed without a timeout.

The small exact-label result is not evidence that only 125 relevant cases exist. In the food and drug agency pool, the stored cause field was the empty string in 6,774 matching records. That observation concerns this query pool; it is not a claim about every administrative record. It does show why a procurement specification based solely on one cause label can miss substantial text that merits review.

The broader market regulator pool is equally easy to misread. Its subject matter may extend beyond food and drugs, and institutional mentions do not prove that the named agency was the defendant. A clean product-specific slice needs reviewed inclusion rules. The gap between these counts is a reason to inspect records, not a reason to advertise the largest number as relevant coverage.

Agency names and dates need their own coverage check

An English query such as “Chinese food regulator” may map to several Chinese names in the documents. A pilot should test historical food and drug administration names, market supervision names and drug supervision names, retaining the name actually used in each record. Normalizing everything into a single English agency label loses information that a researcher may need.

The same caution applies to dates. We aggregated the stored year without adding a year filter. The food and drug agency pool contained 3,408 records in the 2018 bucket and seven in 2025; the market agency pool contained 15,860 in 2020 and 1,439 in 2025. These are query-specific availability counts. They do not establish a change in enforcement activity, business compliance or willingness to litigate.

A stored year may reflect the dataset’s date parsing rather than the date of the underlying inspection or penalty. For a current-law assistant, review the actual document dates and the dates of the events discussed. See our coverage-by-year guide for the wider distinction between document availability and litigation activity.

Preserve what the court decided about each action

Food and drug records can contain several actions that should be modeled separately: a seizure, a penalty, a reconsideration decision and a request concerning the seized goods. A document-level positive or negative label can flatten the distinction that matters most to counsel.

The SPC’s August 18, 2025 collection provides an example. In Case One, a Chinese medicinal materials business challenged a market regulator’s extended seizure of goods. The court confirmed that the prolonged seizure was unlawful and required remedial measures, while declining to order the goods’ immediate return because of drug safety concerns. The appeal was dismissed. Source: SPC business-related administrative coercion typical cases, Case One.

For a research assistant, “the seizure was unlawful” and “the goods must be returned” are different propositions. For an evaluation set, the answer should distinguish them and point to the relevant passages. This is an inference about annotation design from the published example, not an estimate of how other courts resolve similar disputes.

A suitable schema would record the challenged act, the requested remedy, the court’s treatment of each request and any unresolved issue separately. It would also distinguish the present court’s reasoning from a party’s allegation or a quoted earlier decision. These annotations require text review; they should not be assumed to exist merely because a supplier offers structured metadata.

Specify native fields and derived annotations separately

The index used for this check stores document text and fields including title, case number, court, date, year, cause, case type, procedure, province and ingestion dataset. Field availability varies by record. The full index does not contain a ready-made, validated field for every regulatory issue a buyer might want to filter.

Do not assume that each record has a working official judgment URL. Case numbers, court names and dates are useful reference coordinates, but they do not guarantee that an original remains publicly accessible. The official links in this article support its examples; they are not a source-link guarantee for the commercial corpus.

Build an enterprise pilot around reviewed examples

A practical pilot starts with questions your product will actually receive. One might concern a food sampling dispute; another, drug manufacturing classification; another, a seizure or a regulator’s response to a complaint. Define the expected document track for each question before testing retrieval.

1. Agree on the slice. List products, agency-name variants, actions, regions and date boundaries. State whether court enforcement applications and reconsideration-related judgments are included.
2. Inspect candidate records. Check both obvious matches and tempting false matches, including civil judgments quoting regulatory materials. Keep inclusion and exclusion reasons.
3. Evaluate document quality. Examine usable Chinese text, case-number variants, repeated records and separate appeal documents. Report record and document counts with their definitions.
4. Test the English layer. Compare the retrieved Chinese text with the English summary. Require the summary to preserve the challenged act, requested relief and procedural posture.
5. Agree on annotation requirements. Separate available metadata from any new extraction, human review or custom taxonomy. Put acceptance criteria in the delivery specification.

If the pilot reports a sampled relevance rate, its sample must cover the contributing ingestion batches rather than the first search results. Save the seed, sample sizes, weighting and exclusions, and inspect whether one repeated litigant or dispute dominates the sample. Without that evidence, a precise percentage can give a misleading impression of quality.

Do not derive litigation outcome rates from these retrieval pools. Topic membership, document duplication, procedural stage and the remedy under review are unresolved at that point. A research product can still be valuable without turning a collection of heterogeneous records into a claim about likely success.

Delivery options for legal AI and regulatory research

SinoVerdict offers access to 160M+ Chinese court records through enterprise bulk delivery, REST API and MCP, with English queries and English summaries. The headline number describes the overall record corpus. It is not the count of food and drug administrative cases, distinct disputes or records with usable full text.

Bulk delivery may suit a team building its own index and annotations; API access may suit a retrieval pilot; MCP may suit evaluation inside a compatible assistant. The useful choice depends on your integration and the agreed fields, rights and delivery scope. Our delivery overview describes the available access paths.

Before contracting, request a coverage report for the exact regulatory slice. It should make the candidate-query boundaries, date distribution, document availability and duplicate treatment visible. Discuss your intended uses and any derivative or redistribution requirements in the agreement rather than inferring permissions from a dataset’s public description.